Two stage GNSS outlier detection for factor graph optimization based GNSS-RTK/INS/odometer fusion

summary

The gist

Reliable GNSS positioning in complex environments remains a critical challenge due to non-line-ofsight (NLOS) propagation, multipath effects, and frequent signal blockages.

In short

The system uses a two-stage method to detect and reject bad GNSS measurements within a tightly coupled RTK/INS/odometer framework optimized by factor graphs. Stage one uses Doppler data for gross error detection, and stage two uses predicted measurements from the IMU and odometer to refine outlier rejection. This combination significantly improves accuracy in challenging environments.

Key concepts

Two-Stage Outlier Detection Framework
This method employs a hierarchical approach to identify bad GNSS data. The first stage checks for large, gross errors using Doppler measurements alone. The second, more sophisticated stage uses predictions from the integrated IMU and odometer to generate double-difference values, which helps pinpoint and reject remaining outliers.
GNSS-Only Outlier Detection (Doppler)
This initial step detects major pseudo-range errors in GNSS data without other sensors. It compares the temporal change in pseudo-ranges against the rate predicted by Doppler measurements. If they don't match, it signals that the pseudo-range is likely contaminated by large, gross errors.
IMU and Odometer Aided Detection
The second stage leverages pre-integrated IMU and odometer data to predict what the GNSS measurements should look like. By comparing these predictions (double-difference pseudo-range) against the actual GNSS measurements, the system can identify and downweight outliers that are consistent with kinematic inconsistencies.
Factor Graph Optimization (FGO)
FGO is a mathematical framework used to jointly estimate navigation states and sensor parameters. It incorporates constraints from all sensors—GNSS, RTK, INS, and odometry—to find the most accurate overall solution, ensuring robust state estimation even when some measurements are faulty.

Terminology used across episodes

This episode discusses

The paper

Two stage GNSS outlier detection for factor graph optimization based GNSS-RTK/INS/odometer fusion · Read on arXiv

Baoshan Song, Penggao Yan, Xiao Xia, Yihan Zhong, Weisong Wen, Li-Ta Hsu

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: We're looking at the paper: "Two stage GNSS outlier detection for factor graph optimization based GNSS-RTK/INS/odometer fusion".

Title and authors: Rosa: So we're looking at this paper called "Two stage GNSS outlier detection for factor graph optimization based GNSS-RTK/INS/odometer fusion". It tackles that big problem of keeping reliable positioning when the signals are messy, like in a city canyon where reflections and blockages happen a lot.

Dev: Yeah, it's about how to make sure the measurements from the GNSS don't mess up the whole system when you couple it with your inertial navigation and wheel odometry. It sets up this two-stage approach using factor graph optimization, or FGO.

Taro: I mean, I always think about those raw pseudo-range measurements getting completely corrupted by multipath effects when you're moving through complex structures, and this paper seems to be trying to handle that directly in the measurement stream.

Rosa: Exactly. The core idea is they use a two-stage method to clean up those bad measurements before they even get into the main navigation estimate. It’s about being robust against both gross errors in distance and bad quality signals from the satellites themselves.

Dev: The first stage focuses on GNSS only, using Doppler measurements to spot potential issues right away because Doppler is less affected by multipath than the raw pseudo-range values.

Taro: So they’re using the rate of change of range from Doppler to check if the actual measured distance makes sense according to what we expect kinematically. That sounds like a good first filter for sudden, large errors.

Rosa: Right. The second stage gets more sophisticated by bringing in the integrated inertial and odometer data to predict what the GNSS measurement should look like before comparing it to what we actually got from the receiver.

Dev: They generate these predicted double difference pseudo-range measurements using the pre-integrated IMU and odometer data, which lets them catch errors that are more subtle than just gross distance mistakes.

Taro: That sounds like a way to enforce kinematic consistency across all the sensors at once, rather than just checking individual numbers in isolation. How does this prediction actually help when the signals are really degraded?

Rosa: Well, by comparing that predicted double difference measurement with the single difference double difference measurement from the GNSS receiver, they can flag any deviation that suggests an outlier needs to be downweighted in the factor graph optimization.

Title and authors: Dev: That comparison value is what flags the bad data; if it crosses a certain threshold, they mark it as an outlier and reduce its influence on the final position estimate. This whole process happens within the factor graph optimization framework.

Taro: So, when you look at their experiments, what kind of improvement are we actually seeing in real-world conditions?

Rosa: They tested this in a deep urban canyon test, and they saw a significant drop in the Root Mean Square error for the RTK inertial system fusion. It went from zero point five two meters down to zero point three zero meters, which is a forty-two point three percent improvement over what they were comparing it against.

Dev: That forty-two point three percent reduction in error is substantial when you're dealing with these challenging environments where the GNSS signal quality fluctuates so much. It shows the method works better than just using standard pre-processing techniques alone.

Taro: So, if I were driving through a dense urban area and my phone signal starts flickering or getting weird reflections, this paper suggests a way to keep the navigation stable by checking both the rate of change from Doppler and comparing against predictions from my other sensors.

Rosa: That’s it. It gives you two layers of defense: one for quick gross errors using Doppler, and another for refined consistency using your IMU and wheels before it messes with the main position calculation in the factor graph optimization based RTK inertial system fusion.

Dev: The paper shows that this combination helps maintain high positioning accuracy even when the GNSS signals are degraded, which is what we really need for a robust real-time loop.

Taro: For my perspective, it’s interesting how they structured the detection hierarchically; using Doppler first and then building up to the more complex prediction from IMU and odometer data sounds like a very sensible way to handle failure modes.

Rosa: It does feel sensible because it guards against both the big, obvious errors that Doppler catches and the more subtle, geometry-based inconsistencies that only come from fusing all those different sensor inputs.

Dev: And that comparison with the predicted double difference measurement is key because it forces a check on geometric and kinematic consistency across the entire coupled system.

Taro: So what's left for us to consider? They mentioned they focused on GNSS-only outlier mitigation after signal processing, so how does this method integrate into a full real-time control loop where you need to react instantly?

Title and authors: Rosa: That’s where my field work comes in; the paper shows the framework is built for practical deployment in autonomous ground vehicles, meaning it’s designed to be applied within those tight constraints of real-time systems.

Dev: The computational cost of these two stages needs to be manageable for a real-time loop, and I'm interested in how their pre-integration methods keep the latency low during the actual optimization step.

Taro: I wonder about the limitations they admit; they noted that this approach is based on GNSS measurements after signal processing, which means if those initial processing steps are fundamentally flawed, the detection stage might not catch everything.

Rosa: That's a fair point; their own limitation is that it focuses on measurements *after* some initial signal processing has already occurred, so the robustness of the entire two-stage system depends heavily on how well those first steps handle NLOS and multipath before we even start detecting outliers.

Dev: So, they aren't claiming perfect immunity to everything; they are claiming a much better way to handle the outliers that remain after standard processing has done its job, which is a realistic goal for this kind of navigation system.

Taro: For anyone listening who just wants the practical impact, it means that for autonomous systems operating in tricky urban or complex environments, you can expect a much more stable and accurate position estimate than you could get with simpler outlier rejection methods.

Rosa: It's about moving from simply hoping the GNSS data is good enough to actively verifying and weighting every piece of data based on how well it fits the predictions from your other sensors.

Dev: The paper, "Two stage GNSS outlier detection for factor graph optimization based GNSS-RTK/INS/odometer fusion," gives us a concrete method to improve the reliability of tightly coupled navigation systems by using a dual-stage filtering process within a factor graph optimization framework.

Taro: It’s about making sure that when you're relying on the combination of GNSS, IMU, and odometry for your position, you have a structured way to identify and reject the noisy data points that are introducing errors.

Rosa: That’s it. This work shows a specific way to handle those messy real-world conditions where signal quality is not constant, giving us a more consistent positioning result when we need it most.

The paper's summary: Rosa: So, to recap, this paper is about taking that messy data coming from GNSS receivers and cleaning it up in two distinct steps before you feed it into your main navigation engine which is doing factor graph optimization with inertial sensors and wheels.

Dev: Exactly. It’s a two-stage outlier detection framework specifically designed for those tough real-world scenarios where signals are bad, like in a city canyon or when you have lots of reflections.

Taro: I mean, the core idea is using Doppler measurements first to catch gross errors right away because Doppler isn't as sensitive to multipath as the raw pseudo-range distance.

Rosa: Right. That’s stage one, detecting those big distance mistakes by checking if the rate of change you get from Doppler matches what you expect kinematically, and it guards against those huge pseudo-range errors immediately.

Dev: Then stage two kicks in where they use the integrated inertial and odometer data to generate predicted measurements, like these double difference pseudo-range values, which lets them spot more subtle outliers that just don't match the geometry.

Taro: So it’s a hierarchy; you have a quick filter for obvious mistakes with Doppler, and then you have a more detailed check using the integrated system to catch anything else that looks inconsistent.

Rosa: And when they apply this to real tests, like those deep urban canyon trials, they showed a big improvement in accuracy; the error drops from zero point five two meters down to zero point three zero meters, which is a forty-two point three percent reduction.

Dev: That forty-two point three percent drop is pretty solid when you’re dealing with these challenging environments; it shows that this two-stage method actually makes a difference in how accurate the final position estimate is.

Taro: It proves that by combining the fast, low-level check with a more complex system prediction, you can keep the entire navigation system stable even when the GNSS signal quality is really fluctuating.

Rosa: Exactly. It shifts things from just hoping your GNSS data is good enough to actively verifying and downweighting every piece of data based on how well it fits what your other sensors are telling you to expect.

Dev: The paper lays out the whole structure within the factor graph optimization framework, showing how these two detection stages feed their results into the main estimation process for better consistency.

Taro: But they also mention a limitation; this method works best after some initial signal processing has already happened, so if that first step is really bad at handling multipath, this entire detection system might not be able to catch every kind of error.

Rosa: That’s the caveat. It’s not a perfect solution for every single possible signal degradation scenario; it depends on how well your initial processing handles those initial messy inputs before you even start looking for outliers.

Dev: So, while the results are impressive in these tests, we still need to figure out if this two-stage approach is robust enough for the absolute worst-case conditions in a real autonomous system deployment.

Taro: That’s the next step; extending this framework to handle multi-constellation signals or testing it under even more extreme urban canyon conditions would be important for showing its long-term reliability.

The paper's improvements: Taro: So, to summarize where we are now, this paper gives us two specific ways to make sure that when you’re fusing GNSS RTK data with your inertial sensors and wheels into a factor graph optimization system, you aren't letting bad data skew the whole position estimate.

Rosa: Right. The main improvement is that it moves beyond just throwing out obviously bad measurements; it builds a detection process where the first stage catches big errors using Doppler, and the second stage uses predictions from your sensors to spot subtler inconsistencies.

Dev: That refined rejection based on predicted double difference measurements is what really tightens up the safety case for real-time loops because it checks geometric and kinematic consistency across all inputs simultaneously.

Taro: It’s about graceful recovery; instead of the system failing when a signal gets bad, it flags that specific measurement as an outlier and downweights its influence in the optimization factor graph.

Rosa: And I gotta point out the real-world payoff: they tested this in those deep urban canyon trials and saw a forty-two point three percent improvement in accuracy; going from zero point five two meters error down to zero point three zero meters.

Dev: That's a solid number when you’re talking about improving navigation stability in noisy environments, showing that this dual-stage detection actually works better than simpler outlier rejection methods alone.

Taro: It means for someone running an autonomous system, it gives you more confidence in the position estimate because you have a structured way to verify every measurement based on its fit with the rest of your coupled sensor data.

Rosa: It changes how we think about noise; instead of seeing all GNSS errors as random spikes, this method helps us categorize those errors into gross distance mistakes and subtle geometric mismatches that need different kinds of correction.

Dev: The paper shows that you can maintain high positioning accuracy even when the GNSS signal quality is fluctuating wildly, which is a big deal for any application requiring real-time performance.

Taro: So, the implication here for autonomy research is that we can start building systems where the navigation pipeline itself has this built-in self-checking mechanism to handle unexpected environmental noise gracefully.

Rosa: Exactly. It’s not just about getting a better number; it’s about building a more robust system that can operate reliably when the world misbehaves, which is what we always talk about out here in the field.

Dev: The paper also lays out the path forward for future work, focusing on extending this to handle multi-constellation GNSS signals and testing it under even more extreme urban canyon conditions.

Taro: That makes sense; pushing it to those harder conditions is where you really find out if this framework has the endurance to survive in truly messy environments.

Conclusion: Rosa: So to wrap up, this paper, "Two stage GNSS outlier detection for factor graph optimization based GNSS RTK/INS/odometer fusion," shows a structured way to clean up messy satellite measurements using two filtering stages before they get into your main navigation estimate.

Dev: It really solidifies the safety case for tightly coupled systems by providing a dual-layered defense against both gross distance errors and subtle geometric inconsistencies in real-time loops.

Taro: I just think it’s important that we see this level of structured filtering applied directly to the factor graph optimization framework, because it shows how you can make the system more resilient when things go wrong in tough environments.

Rosa: Yeah, so what this means for us is that we move from just hoping the GNSS data is good enough to actively verifying and weighting every piece of data based on how well it fits the predictions from your inertial sensors and wheels.

Dev: It gives us a more predictable error profile, which is crucial for designing robust real-time control systems that need to maintain high accuracy even when signals are degraded.

Taro: I'm interested in what this suggests for future work, like extending this to handle multi-constellation signals or testing it under even more extreme urban canyon conditions where the signal quality really drops off.

Rosa: Right, so they’re looking at those bigger challenges next, which is exactly where we need to take this kind of research.

Dev: We'll see how they tackle those extensions because right now, the core two-stage mechanism seems very promising for improving consistency in complex navigation scenarios.

More episodes

← Home